Neural networks are increasingly deployed on mobile and IoT devices, offering low latency and better privacy than cloud solutions. Yet models embedded in apps can be reverse‑engineered and exposed to attackers, enabling cross‑node attacks: adversarial examples crafted on a compromised device often transfer to other nodes that share the same model.
To mitigate this risk, we introduce VeilNN, a defense framework that reduces adversarial transferability by injecting controlled model diversity while preserving the original task performance. VeilNN applies stealthy, irreversible obfuscation at both architectural and parameter levels through network cell obfuscation and layer assembling, and works with MLPs, CNNs, and Vision Transformers alike.
Extensive experiments evaluate VeilNN against four representative adversarial attacks. The average defense success rate reaches 49.19% and peaks at 81.19%, representing more than a 33.16% improvement over existing defenses. These findings demonstrate that modest model diversification can substantially weaken cross‑device adversarial transfer.
Review: VeilNN offers a practical approach to securing edge AI by structurally obfuscating models without sacrificing functionality, opening a promising direction for robust deployment on heterogeneous devices.